Богдан

Bohdan Kolisnyk
13, Ivasyuka Avenue, Kyiv, Ukraine 04210
[відкрити контакти](див. вище в блоці «контактна інформація») # [відкрити контакти](див. вище в блоці «контактна інформація») ï [відкрити контакти](див. вище в блоці «контактна інформація») kaggle.com/bohdankolisnyk
Junior DL/ML engineer with hands-on experience in audio and bioacoustic classification (Kaggle BirdCLEF: embeddings,
class-imbalance handling, ensembling, ONNX deployment) and signal-driven computer vision (UAV/object detection on noisy,
real-world sensor data). Experience in developing from raw signal to production-ready inference.
Education
Kyiv School of Economics Sep. 2023 – May 2027
Economics and Big Data, 4th Year Kyiv, Ukraine
Relevant Coursework
• Machine Learning • NLP • DE Concepts • Audio Deep Learning
• Deep Learning • Computer Vision • Gen AI in CV • Reinforcement Learning
Experience
Piraeus Bank April 2026 – July 2026
Data Science Intern Kyiv, Ukraine
• Built complex Oracle SQL queries (CTEs, JSON TABLE) for NBU regulatory reporting, diagnosing and resolving

cross-schema performance bottlenecks via database link query optimization.
• Designed Power BI data models, DAX measures, and Power Query (M) pipelines connecting to Oracle via ODBC,

migrating legacy Excel reports into BI dashboards.
• Researched Grafana/Prometheus integration with Power BI for unifying corporate BI tooling.

BBS Insurance May 2025 – July 2025
Data Analyst Intern Kyiv, Ukraine
• Analyzed financial reporting and operational metrics, optimizing reporting workflows and improving data accuracy.

• Designed and maintained automated dashboards for the analytics department, supporting data-driven decision-making.

Competitions and Researches
BirdCLEF+ 2026, Kaggle | Team Member May 2026
• Built a bioacoustic species classification pipeline using Perch v2 audio embeddings and MLP heads, achieving CV AUC

∼0.98 via stratified 5-fold cross-validation.
• Improved generalization to the leaderboard distribution by incorporating soundscape-labeled training data and EMA

fine-tuning, then built a final 8-model ONNX ensemble with TTA, raising the leaderboard score from ∼0.81 to 0.896+.
Long-Term Tracking of Military Vehicles in Low-FPS UAV Footage | Co-Author and Researcher May 2026
• Built a multi-object tracking pipeline (YOLOv8s detector + MASA-R50 tracker with ResNet50 backbone, FPN, and

Hungarian matching) to track military vehicles in low-FPS UAV footage on a custom-curated dataset.
• Engineered a per-track gallery memory mechanism improving HOTA by 4.6% (GT mode) and 5.9% (end-to-end) while

cutting ID switches up to 4.6x; diagnosed detection quality as the main bottleneck via correlation analysis and applied
contrastive fine-tuning (InfoNCE) to further reduce switches.
Temporal Layers in CNN-Based UAV Detection | Co-Author and Researcher May – July 2025
• Co-authored a research paper benchmarking four CNN backbones (MobileNetV2, EfficientNet, ResNet50, ConvNeXt)

against four temporal encoders (ConvLSTM, 3D convolution, optical flow, transformer) for UAV detection on infrared
video, evaluated on the Anti-UAV300 dataset.
• Built the MobileNetV2 + ConvLSTM pipeline end-to-end: designed a temporal encoder that stacks 3 consecutive frames

into a 9-channel hidden-state feature map, modified the backbone’s input layer, and implemented a custom multi-task
detection head - classification + bounding-box regression.
• Evaluated the pipeline on standard and hard validation subsets using [email protected], precision, recall, and IoU, contributing to

the paper’s core finding that temporal-layer benefit is backbone-dependent rather than universally positive.
Technical Skills
Languages: Python, R, SQL, M
Developer Tools: VS Code, Git, Apache Airflow, Docker, GitHub, Colab, MiniConda, Bash
Frameworks/Libraries: PyTorch, Torchaudio, Librosa, Scikit-learn, ONNX, OpenCV, Pandas, NumPy
Extracurricular
NLP Summer School June 2026
Participant Kyiv School of Economics
• Completed a 4-day intensive program (KSE x Superhuman) on the modern LLM/NLP stack — transformer alternatives,

agentic harnesses, RAG, synthetic data generation, AI safety, and LLM distillation/MLOps — with sessions led by a
Kaggle Competitions Grandmaster and researchers from the KSE NLP Lab, SoftServe, and Superhuman.